Critical Nvidia Triton server flaws exposed by Wiz

Cloud security firm Wiz disclosed three critical vulnerabilities (CVE-2025-23319, CVE-2025-23320, CVE-2025-23334) in Nvidia’s Triton Inference Server, an open-source platform for deploying AI models. The flaws reside in the Python backend on Windows and Linux, allowing remote, unauthenticated attackers to execute arbitrary code, trigger denial-of-service, or leak sensitive data. By chaining a minor information disclosure into a full system compromise, adversaries could steal AI models, manipulate responses, and gain deeper network access. Nvidia has addressed these issues in Triton version 25.07, and users are urged to update immediately to secure their AI deployments.

Read more:

https://www.securityweek.com/nvidia-triton-vulnerabilities-pose-big-risk-to-ai-models/